Container bloat: Runner image 65GB due to incorrect build context

[jeremyeder]

Summary

The vteam_claude_runner container image was bloated to 65GB in production (3.7GB locally). Root cause: incorrect Docker build context copying development artifacts including .venv directories (432MB) and unrelated components.

Fix: Added .dockerignore, corrected build context path, and updated COPY command.

Impact:

• Build context: 413MB → 70MB (83% reduction)
• Image size: 3.7GB → 2.8GB (25%+ reduction, potentially 95% in CI/CD)

Root Cause

The build context was set to ./components/runners/ (parent directory) instead of ./components/runners/claude-code-runner/, causing Docker to copy:

• 3 virtual environments (.venv, .venv-langfuse, venv/) - 432 MB
  • claude_agent_sdk/_bundled/claude - 173 MB
  • macOS binaries (.so files) - useless in Linux containers
  • Multiple ruff binaries - 22-29 MB each
• mcp-ambient-server/ subdirectory - 109 MB (unrelated component)
• Test files, lock files, and other dev artifacts

📊 Build Context Analysis (from diagnosis script)

Current (incorrect) build context: components/runners/
541M    total

Top files copied:
173M    .venv/lib/python3.13/site-packages/claude_agent_sdk/_bundled/claude
29M     mcp-ambient-server/.venv/bin/ruff
22M     .venv/bin/ruff
20M     .venv/lib/python3.13/site-packages/cryptography/hazmat/bindings/_rust.abi3.so
20M     mcp-ambient-server/venv/lib/python3.14/site-packages/cryptography/...

Build context transfer:

• Current (incorrect): 413.98 MB
• Fixed (correct): 70.86 MB

Changes Made

1. Added .dockerignore

File: components/runners/claude-code-runner/.dockerignore

Excludes:

• Virtual environments (.venv/, .venv-*/, venv/)
• Tests and coverage files
• Lock files (uv.lock)
• IDE files, logs, docs

2. Fixed Dockerfile

File: components/runners/claude-code-runner/Dockerfile (line 64)

- COPY claude-code-runner /app/claude-runner
+ COPY . /app/claude-runner

3. Updated GitHub Actions Workflow

File: .github/workflows/components-build-deploy.yml (line 102)

  - name: claude-code-runner
-   context: ./components/runners
+   context: ./components/runners/claude-code-runner
    image: quay.io/ambient_code/vteam_claude_runner
    dockerfile: ./components/runners/claude-code-runner/Dockerfile

4. Updated Makefile

File: Makefile (line 138-142)

  build-runner:
-   @cd components/runners && $(CONTAINER_ENGINE) build ... -f claude-code-runner/Dockerfile .
+   @$(CONTAINER_ENGINE) build ... -f components/runners/claude-code-runner/Dockerfile \
+       components/runners/claude-code-runner

Expected Results

After this fix:

• ✅ Build context: ~70 MB (vs 400+ MB)
• ✅ Image size: 2.5-2.8 GB (vs 65 GB or 3.7+ GB)
• ✅ Faster builds (less context transfer)
• ✅ No dev artifacts in production images

The 2.8 GB final size is normal for this stack:

• Base image (UBI9 Python 3.11): ~600 MB
• System packages (git, gh, Node.js): ~200 MB
• Python deps (160 packages): ~1.5 GB (anthropic[vertex], langfuse, fastapi, etc.)

Testing

A diagnosis script was created to reproduce and verify the fix locally. See attached files for:

• Full diagnosis report
• Build logs (before/after)
• Layer-by-layer comparison

🔍 How to verify the fix locally

cd ~/repos/platform

# Check build context size
tar -czf - -C components/runners/claude-code-runner . | wc -c | numfmt --to=iec-i

# Should show ~70-120 MB (vs 150+ MB before)

# Build image
docker build \
    -t vteam_claude_runner:test \
    -f components/runners/claude-code-runner/Dockerfile \
    components/runners/claude-code-runner/

# Check size
docker images vteam_claude_runner:test
# Should show ~2.5-2.8 GB

Action Items

• Add .dockerignore file
• Update Dockerfile COPY command
• Update GitHub Actions workflow context
• Update Makefile build context
• Monitor next CI/CD build for size verification
• Update documentation if needed

Related

Commit: (will be added after PR)

---

cc: @jeder

[jeremyeder]

Full Diagnosis Report

Attached is the complete diagnostic report from the bloat analysis script that reproduced and identified the issue locally.

Key metrics from diagnosis:

• Current (incorrect) build: 3.73 GB image, 413 MB context transferred
• Fixed build: 2.78 GB image, 70 MB context transferred
• Savings: 950 MB image size, 343 MB context size

The diagnosis script builds both configurations and compares them layer-by-layer to identify exactly where the bloat occurs.

See the full report attached in the artifacts folder of the fix PR.